OpenAI launches GPT-6 Sol and Luna, halving API prices
Sol codes at $2/$10, Luna clerks at $0.10/$0.50, both with a 272K price cliff — plus an unpatched Meta Muse Mac backdoor and AMD's open long-context hybrids.
Copy markdown
Sol: frontier coding at half the price
GPT-6 Sol runs $2/M in and $10/M out — half of GPT-5.6's $4/$20 — and is tuned for complex work like coding, where OpenAI says it makes about half as many mistakes as before. It's live now in Codex, ChatGPT Work, and the API, so your top-tier coding-agent bill roughly halves overnight.
Luna: the $0.10 workhorse tier
Luna handles high-volume, goal-clear jobs — summarizing, extracting, quick answers — at $0.10/M in and $0.50/M out, down from $0.20/$1.20. It's in the API plus the desktop app and Free/Go tiers, making batch and RAG pipelines roughly 2x cheaper to run.
Mind the 272K token price cliff
Both models take up to 1.05M input / 128K output tokens, but any request over 272K input tokens bills at double rate. The giant context is real — just do not let an agent silently blow past 272K and 2x your bill.
The 'beats Opus' claim, checked
OpenAI says Sol and Luna outdo Anthropic’s Fable and Opus, but independent runs show last-gen GPT-5.6 Sol still edges the new Sol on DeepSWE (72.7% vs 68.8%). The real win is cost — about $2.74 vs $6.46 per task — not raw score, so benchmark your own workload before switching.
Meta Muse has an unpatched Mac backdoor
Researcher Patrick Wardle disclosed a 0-day: an undocumented dictation-endpoint setting lets any local app on your Mac redirect Muse, inject instructions it trusts, and steal your session token to drive the assistant across devices (he pivoted to an iPhone’s location). No patch yet — mitigate by uninstalling or disabling voice input.
AMD open-sources 14 long-context hybrids
AMD's Zebra-HyLo drops 14 Apache-2.0 checkpoints (built on Llama-3.2 and Qwen3) that cut KV-cache 92-98% and served 2M tokens on eight MI300X GPUs. You can 'upcycle' an existing Transformer into a long-context hybrid with ~10B tokens of post-training instead of pretraining — though the 64K variants use non-commercial data.